Machine Learning Model Deployment Using Real-Time Physiological Monitoring: Use Case of Detecting Delayed Cerebral Ischemia

Abstract

Deploying machine learning models in a clinical setting is challenging. Here we demonstrated a modular model deployment pipeline for for ContinuOuS Monitoring tool for delayed cerebral IsChemia (COSMIC) that was successfully implemented on a trained and validated temporal machine learning algorithm that detects delayed cerebral ischemia after subarachnoid hemorrhage. The pipeline was able to ingest demographic data and near real-time continuous physiological data, run through a model, and output the results twice a day automatically. It was a highly collaborative effort among clinical neurologists, the research team, and the IT innovation team. We illustrate the technical run time challenges and mitigations in each of the three components of the pipeline: data, model (modular), and output communication. Future work is user-centered participatory design and rapid agile prototyping of effective model output communication and clinical trial of efficacy in order to understand how the clinical decision support tool performs and is adopted in a real clinical setting.